detection of skin cancer - dtu
TRANSCRIPT
1
Detection of skin cancerJan LarsenSection for Cognitive SystemsDTU Informatics
isp.imm.dtu.dk
1
p
Systems neuro-science
MultimediaDigital economy
Machine learning Signal
processingCognitive modeling
Biomedical
Demining and tools Monitor
systems
Mobile services
y
06/11/20092 DTU Informatics, Technical University of Denmark
for EODHCI
systems
extraction of meaningful and actionable information by ubiquitous learning from data
2
Systems neuro-science
MultimediaDigital economy
Biomedical
•Neuroimaging(PET EEG fMRI)
www.intelligentsound.org
Machine learning Signal
processingCognitive modeling
Biomedical
Demining and tools Monitor
systems
Mobile services
y(PET,EEG,fMRI)
•EEG sensor for early warning of low blood suguar
•Improved SP in hearing aids
•Cognitive modeling
www.cimbi.org
hendrix.imm.dtu.dk
isp.imm.dtu.dk
06/11/20093 DTU Informatics, Technical University of Denmark
for EODHCI
systems
extraction of meaningful and actionable information by ubiquitous
learning from data
•Cognitive modeling
Skin cancer •More than 800 cases in Denmark yearly
B i i•Annual increase 5-10%
•Benign nevi
•Atypical nevi
•Malignant melanoma •Inexperienced
doctors detect 31%
06/11/20094 DTU Informatics, Technical University of Denmark
•Experienced doctors detect 63-75%
3
Skin cancer •More than 800 cases in Denmark yearly
B i i•Annual increase 5-10%
•Benign nevi
•Atypical nevi
•Malignant melanoma •Inexperienced
doctors detect 31%
06/11/20095 DTU Informatics, Technical University of Denmark
•Experienced doctors detect 63-75%
Objectives• Develop a cost-effective and practical tool for diagnosis support• Gain more insight into the understanding of factors in the development of
skin cancerskin cancer
06/11/20096 DTU Informatics, Technical University of Denmark
4
Cross-disciplinary research
Signal and Image processing
StatisticsMachine learning
Domain knowledge
06/11/20097 DTU Informatics, Technical University of Denmark
StatisticsMachine learning
Outline• Machine learning framework for skin cancer detection
– Involves all issues of machine learning• An image processing system for skin cancer detection• An image processing system for skin cancer detection
– Involves feature selection, projection and integration– Involves linear and nonlinear classifiers
• Other approaches• Summary
06/11/20098 DTU Informatics, Technical University of Denmark
5
The potential of learning machines• Most real world problems are too complex to be handled by classical
physical models• In most real world situations there is access to data describing properties • In most real world situations there is access to data describing properties
of the problem• Learning machines can offer
– Learning of optimal prediction/decision/action– Adaptation to the usage environment– New insights into the problem and suggestions for improvement
06/11/20099 DTU Informatics, Technical University of Denmark
A short history of learning machines
clas
sica
l
mod
ern
ADALINE
Neural nets
Gaussian processes
Kernel machines
06/11/200910 DTU Informatics, Technical University of Denmark
Gaussian processesMixture of experts
6
Issues in machine learning
Data•quantity
•stationarity
•quality
•structure
Features
•representation
•selection
•extraction
•integration
Models
•structure
•type
•learning
•selection and
integration
Evaluation
•performance
•robustness
•complexity
•interpretation and visualization
•HCI
06/11/200911 DTU Informatics, Technical University of Denmark
•HCI
Issues in machine learning•parametric: linear, nonlinear, mixture models
•non-Data•quantity
•stationarity
•quality
•structure
Features
•representation
•selection
•extraction
•integration
Models
•structure
•type
•learning
•selection and
integration
•unsupervised
•semi-supervised
•supervised
f i
Evaluation
•performance
•robustness
•complexity
•interpretation and visualization
•HCI
parametric: kernel, Gaussian processes, clustering
•noise models
•integration of prior and domain knowledge
06/11/200912 DTU Informatics, Technical University of Denmark
•cost function
•maximum likelihood
•Bayesian
•online vs. off-line
•HCIknowledge
7
Dermatoscopy imaging technique
06/11/200913 DTU Informatics, Technical University of Denmark
Domain knowledge – dematoscopic features
06/11/200914 DTU Informatics, Technical University of Denmark
8
Feature extraction
06/11/200915 DTU Informatics, Technical University of Denmark
Median filtering
06/11/200916 DTU Informatics, Technical University of Denmark
Removal of impulsive noise
9
Feature extraction
06/11/200917 DTU Informatics, Technical University of Denmark
Segmentation
06/11/200918 DTU Informatics, Technical University of Denmark
10
Feature extraction
06/11/200919 DTU Informatics, Technical University of Denmark
Assymetry
06/11/200920 DTU Informatics, Technical University of Denmark
11
Feature extraction
06/11/200921 DTU Informatics, Technical University of Denmark
Edge abruptness
06/11/200922 DTU Informatics, Technical University of Denmark
12
Edge abruptness
06/11/200923 DTU Informatics, Technical University of Denmark
Feature extraction
06/11/200924 DTU Informatics, Technical University of Denmark
13
Color prototypes
06/11/200925 DTU Informatics, Technical University of Denmark
Segmentation into color prototypes
06/11/200926 DTU Informatics, Technical University of Denmark
14
Bayes classifier
06/11/200927 DTU Informatics, Technical University of Denmark
Bayes classifier
06/11/200928 DTU Informatics, Technical University of Denmark
15
Neural network classifier
06/11/200929 DTU Informatics, Technical University of Denmark
Likelihood learning
06/11/200930 DTU Informatics, Technical University of Denmark
Training set: N samples of related x(k) and classes y(k)
16
Generalization• How well are we doing on future data from the same problem?
06/11/200931 DTU Informatics, Technical University of Denmark
Bias Variance dilemma
06/11/200932 DTU Informatics, Technical University of Denmark
17
06/11/200933 DTU Informatics, Technical University of Denmark
Confusion matrix
06/11/200934 DTU Informatics, Technical University of Denmark
18
Other techniques – Raman spectroscopy• A NIR laser beam excites molecules in the skin• The Raman scattering is a frequency shift in the reflected light which is
related to the molecule structurerelated to the molecule structure
06/11/200935 DTU Informatics, Technical University of Denmark
Raman spectrum
•MM: malignant gmelanoma
•NV: pigmented navi
•BCC: basal cell carcinoma
•SK: seborrhoeickeratosis
06/11/200936 DTU Informatics, Technical University of Denmark
keratosis
•NOR: normal
19
Raman classification results
06/11/200937 DTU Informatics, Technical University of Denmark
Ref: Sigurdur Sigurdsson *’s are predicted values using a NN
Further reading• Hintz-Madsen, M., A probabilistic framework for
classification of dermatoscopic images, pp. 156, I f ti d M th ti l M d lli T h i l Informatics and Mathematical Modelling, Technical University of Denmark, DTU, 1998
• Sigurdsson, S., A Probabilistic Framework for Detection of Skin Cancer by Raman Spectra, pp. 202, Informatics and Mathematical Modelling, Technical University of Denmark, DTU, 2003
• Have, A. S., Datamining on distributed medical databases, I f ti d M th ti l M d lli T h i l
06/11/200938 DTU Informatics, Technical University of Denmark
Informatics and Mathematical Modelling, Technical University of Denmark, DTU, 2003
• Papers accessible via http://isp.imm.dtu.dk
20
Related courses• 02450 Introduction to Machine Learning and Data Modeling• 02451 Digital Signal Processing• 02457 Nonlinear Signal Processing• 02457 Nonlinear Signal Processing• 02459 Machine Learning for Signal Processing• 02501 Digital image analysis, vision and computer graphics • 02505 Medical Image Analysis • 31565 Advanced topics in Biomedical Signal Processing
06/11/200939 DTU Informatics, Technical University of Denmark
Summary• Machine learning is, and will become, an important component in most
real world applications• Designing a system involves cross-disciplinary competence – domain • Designing a system involves cross disciplinary competence domain
knowledge, features, classifiers etc.• Automatic detection of skin cancer for diagnosis support is possible
06/11/200940 DTU Informatics, Technical University of Denmark